姬宜欣 发表于 2025-7-31 14:33:04

锚框 anchor box

博客地址:https://www.cnblogs.com/zylyehuo/
参考 《动手学深度学习》第二版
代码总览

# 锚框%matplotlib inline
import torch
from d2l import torch as d2ltorch.set_printoptions(2)# 精简输出精度
def multibox_prior(data, sizes, ratios):
    """生成以每个像素为中心具有不同形状的锚框"""
    in_height, in_width = data.shape[-2:]
    device, num_sizes, num_ratios = data.device, len(sizes), len(ratios)
    boxes_per_pixel = (num_sizes + num_ratios - 1)
    size_tensor = torch.tensor(sizes, device=device)
    ratio_tensor = torch.tensor(ratios, device=device)

    # 为了将锚点移动到像素的中心,需要设置偏移量。
    # 因为一个像素的高为1且宽为1,我们选择偏移我们的中心0.5
    offset_h, offset_w = 0.5, 0.5
    steps_h = 1.0 / in_height# 在y轴上缩放步长
    steps_w = 1.0 / in_width# 在x轴上缩放步长

    # 生成锚框的所有中心点
    center_h = (torch.arange(in_height, device=device) + offset_h) * steps_h
    center_w = (torch.arange(in_width, device=device) + offset_w) * steps_w
    shift_y, shift_x = torch.meshgrid(center_h, center_w, indexing='ij')
    shift_y, shift_x = shift_y.reshape(-1), shift_x.reshape(-1)

    # 生成“boxes_per_pixel”个高和宽,
    # 之后用于创建锚框的四角坐标(xmin,xmax,ymin,ymax)
    w = torch.cat((size_tensor * torch.sqrt(ratio_tensor),
                   sizes * torch.sqrt(ratio_tensor)))\
                   * in_height / in_width# 处理矩形输入
    h = torch.cat((size_tensor / torch.sqrt(ratio_tensor),
                   sizes / torch.sqrt(ratio_tensor)))
    # 除以2来获得半高和半宽
    anchor_manipulations = torch.stack((-w, -h, w, h)).T.repeat(
                                        in_height * in_width, 1) / 2

    # 每个中心点都将有“boxes_per_pixel”个锚框,
    # 所以生成含所有锚框中心的网格,重复了“boxes_per_pixel”次
    out_grid = torch.stack(,
                dim=1).repeat_interleave(boxes_per_pixel, dim=0)
    output = out_grid + anchor_manipulations
    return output.unsqueeze(0)# 返回的锚框变量Y的形状是(批量大小,锚框的数量,4)img = d2l.plt.imread('./assets/catdog.jpg')
h, w = img.shape[:2]
print(h, w)
X = torch.rand(size=(1, 3, h, w))
Y = multibox_prior(X, sizes=, ratios=)
Y.shape
# 访问以(250,250)为中心的第一个锚框boxes = Y.reshape(h, w, 5, 4)
boxes
# 显示以图像中以某个像素为中心的所有锚框def show_bboxes(axes, bboxes, labels=None, colors=None):
    """显示所有边界框"""
    def _make_list(obj, default_values=None):
      if obj is None:
            obj = default_values
      elif not isinstance(obj, (list, tuple)):
            obj =
      return obj

    labels = _make_list(labels)
    colors = _make_list(colors, ['b', 'g', 'r', 'm', 'c'])
    for i, bbox in enumerate(bboxes):
      color = colors
      rect = d2l.bbox_to_rect(bbox.detach().numpy(), color)
      axes.add_patch(rect)
      if labels and len(labels) > i:
            text_color = 'k' if color == 'w' else 'w'
            axes.text(rect.xy, rect.xy, labels,
                      va='center', ha='center', fontsize=9, color=text_color,
                      bbox=dict(facecolor=color, lw=0))# 以(250,250)为中心的锚框d2l.set_figsize()
bbox_scale = torch.tensor((w, h, w, h))
fig = d2l.plt.imshow(img)
show_bboxes(fig.axes, boxes * bbox_scale,
            ['s=0.75, r=1', 's=0.5, r=1', 's=0.25, r=1', 's=0.75, r=2',
             's=0.75, r=0.5'])

# 交并比(IoU)def box_iou(boxes1, boxes2):
    """计算两个锚框或边界框列表中成对的交并比"""
    box_area = lambda boxes: ((boxes[:, 2] - boxes[:, 0]) *
                              (boxes[:, 3] - boxes[:, 1]))
    # boxes1,boxes2,areas1,areas2的形状:
    # boxes1:(boxes1的数量,4),
    # boxes2:(boxes2的数量,4),
    # areas1:(boxes1的数量,),
    # areas2:(boxes2的数量,)
    areas1 = box_area(boxes1)
    areas2 = box_area(boxes2)
    # inter_upperlefts,inter_lowerrights,inters的形状:
    # (boxes1的数量,boxes2的数量,2)
    inter_upperlefts = torch.max(boxes1[:, None, :2], boxes2[:, :2])
    inter_lowerrights = torch.min(boxes1[:, None, 2:], boxes2[:, 2:])
    inters = (inter_lowerrights - inter_upperlefts).clamp(min=0)
    # inter_areasandunion_areas的形状:(boxes1的数量,boxes2的数量)
    inter_areas = inters[:, :, 0] * inters[:, :, 1]
    union_areas = areas1[:, None] + areas2 - inter_areas
    return inter_areas / union_areas# 将真实边界框分配给锚框def assign_anchor_to_bbox(ground_truth, anchors, device, iou_threshold=0.5):
    """将最接近的真实边界框分配给锚框"""
    num_anchors, num_gt_boxes = anchors.shape, ground_truth.shape
    # 位于第i行和第j列的元素x_ij是锚框i和真实边界框j的IoU
    jaccard = box_iou(anchors, ground_truth)
    # 对于每个锚框,分配的真实边界框的张量
    anchors_bbox_map = torch.full((num_anchors,), -1, dtype=torch.long,
                                  device=device)
    # 根据阈值,决定是否分配真实边界框
    max_ious, indices = torch.max(jaccard, dim=1)
    anc_i = torch.nonzero(max_ious >= iou_threshold).reshape(-1)
    box_j = indices
    anchors_bbox_map = box_j
    col_discard = torch.full((num_anchors,), -1)
    row_discard = torch.full((num_gt_boxes,), -1)
    for _ in range(num_gt_boxes):
      max_idx = torch.argmax(jaccard)
      box_idx = (max_idx % num_gt_boxes).long()
      anc_idx = (max_idx / num_gt_boxes).long()
      anchors_bbox_map = box_idx
      jaccard[:, box_idx] = col_discard
      jaccard = row_discard
    return anchors_bbox_map# 标记类别和偏移量def offset_boxes(anchors, assigned_bb, eps=1e-6):
    """对锚框偏移量的转换"""
    c_anc = d2l.box_corner_to_center(anchors)
    c_assigned_bb = d2l.box_corner_to_center(assigned_bb)
    offset_xy = 10 * (c_assigned_bb[:, :2] - c_anc[:, :2]) / c_anc[:, 2:]
    offset_wh = 5 * torch.log(eps + c_assigned_bb[:, 2:] / c_anc[:, 2:])
    offset = torch.cat(, axis=1)
    return offsetdef multibox_target(anchors, labels):
    """使用真实边界框标记锚框"""
    batch_size, anchors = labels.shape, anchors.squeeze(0)
    batch_offset, batch_mask, batch_class_labels = [], [], []
    device, num_anchors = anchors.device, anchors.shape
    for i in range(batch_size):
      label = labels
      anchors_bbox_map = assign_anchor_to_bbox(
            label[:, 1:], anchors, device)
      bbox_mask = ((anchors_bbox_map >= 0).float().unsqueeze(-1)).repeat(
            1, 4)
      # 将类标签和分配的边界框坐标初始化为零
      class_labels = torch.zeros(num_anchors, dtype=torch.long,
                                 device=device)
      assigned_bb = torch.zeros((num_anchors, 4), dtype=torch.float32,
                                  device=device)
      # 使用真实边界框来标记锚框的类别。
      # 如果一个锚框没有被分配,标记其为背景(值为零)
      indices_true = torch.nonzero(anchors_bbox_map >= 0)
      bb_idx = anchors_bbox_map
      class_labels = label.long() + 1
      assigned_bb = label
      # 偏移量转换
      offset = offset_boxes(anchors, assigned_bb) * bbox_mask
      batch_offset.append(offset.reshape(-1))
      batch_mask.append(bbox_mask.reshape(-1))
      batch_class_labels.append(class_labels)
    bbox_offset = torch.stack(batch_offset)
    bbox_mask = torch.stack(batch_mask)
    class_labels = torch.stack(batch_class_labels)
    return (bbox_offset, bbox_mask, class_labels)# 一个例子ground_truth = torch.tensor([,
                         ])
anchors = torch.tensor([, ,
                  , ,
                  ])

fig = d2l.plt.imshow(img)
show_bboxes(fig.axes, ground_truth[:, 1:] * bbox_scale, ['dog', 'cat'], 'k')
show_bboxes(fig.axes, anchors * bbox_scale, ['0', '1', '2', '3', '4']);

# 根据狗和猫的真实边界框,标注这些锚框的分类和偏移量labels = multibox_target(anchors.unsqueeze(dim=0),
                         ground_truth.unsqueeze(dim=0))labels
labels
labels
# 应用逆偏移变换来返回预测的边界框坐标def offset_inverse(anchors, offset_preds):
    """根据带有预测偏移量的锚框来预测边界框"""
    anc = d2l.box_corner_to_center(anchors)
    pred_bbox_xy = (offset_preds[:, :2] * anc[:, 2:] / 10) + anc[:, :2]
    pred_bbox_wh = torch.exp(offset_preds[:, 2:] / 5) * anc[:, 2:]
    pred_bbox = torch.cat((pred_bbox_xy, pred_bbox_wh), axis=1)
    predicted_bbox = d2l.box_center_to_corner(pred_bbox)
    return predicted_bbox# 以下nms函数按降序对置信度进行排序并返回其索引def nms(boxes, scores, iou_threshold):    """对预测边界框的置信度进行排序"""    B = torch.argsort(scores, dim=-1, descending=True)    keep = []# 保留预测边界框的指标    while B.numel() > 0:      i = B      keep.append(i)      if B.numel() == 1: break      iou = box_iou(boxes.reshape(-1, 4),                      boxes, :].reshape(-1, 4)).reshape(-1)      inds = torch.nonzero(iou
页: [1]
查看完整版本: 锚框 anchor box